Accepted Manuscript

A novel version of simulated annealing based on linguistic patterns

for solving facility layout problems

Jerzy Grobelny , Rafał Michalski

PII: S0950-7051(17)30119-3
DOI: 10.1016/j.knosys.2017.03.001
Reference: KNOSYS 3849

To appear in: Knowledge-Based Systems

Received date: 9 November 2016

Revised date: 31 January 2017

Please cite this article as: Jerzy Grobelny , Rafał Michalski , A novel version of simulated annealing
based on linguistic patterns for solving facility layout problems, Knowledge-Based Systems (2017), doi:
10.1016/j.knosys.2017.03.001

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service
to our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please
note that during the production process errors may be discovered which could affect the content, and
all legal disclaimers that apply to the journal pertain.
 ACCEPTED MANUSCRIPT

A novel version of simulated annealing based on linguistic patterns for solving

facility layout problems

Jerzy Grobelny*
Rafał Michalski*

*
Faculty of Computer Science and Management (W8)

T
Wrocław University of Science and Technology,
27 Wybrzeże Wyspiańskiego,

IP
50-370 Wrocław, Poland
e-mail: Jerzy.Grobelny@pwr.edu.pl, Rafal.Michalski@pwr.edu.pl

CR
JerzyGrobelny.com, RafalMichalski.com
phone: +48 71 348 50 50
fax: +48 71 320 34 32

US
2016
AN
Abstract
The paper presents a novel version of the simulated annealing algorithm based on linguistic patterns
(LP) and fuzzy theory approach. The article describes shortly the linguistic patterns and shows on a
M

simple illustrative example the idea of using the Łukasiewicz formula as a criterion for the facility layout
optimization. Next, the detailed description of our LP version of simulated annealing is presented. The
influence of the proposed algorithm parameters on the effectiveness of our approach is then examined
ED

in a simulation experiment involving four different types of facility layout problems. The outcomes
from the first experiments are used for optimally setting the parameters of our proposal in the second
simulation study focused on verifying its effectiveness for objects with uniform and variable sizes. The
obtained results show that the presented procedure, apart from producing decent results in terms of the
PT

classic goal function and the linguistic pattern criterion, provided solutions that were qualitatively
different than those generated by a crisp version.
CE

Keywords
facility layout problem, linguistic patterns, fuzzy sets, simulated annealing, optimization
AC

1
 ACCEPTED MANUSCRIPT

1. Introduction
The facility layout problem (FLP) is commonly related with arranging various types of objects in

possible locations taking into account their connections. Sometimes, additional restrictions regarding

either objects‘ dimensions or their target places are also involved in the optimization process.

Researchers and practitioners come across these types of problems in diverse situations. In such areas

as industrial engineering or logistics, the main goal is to minimize transport costs that are considered to

account from 20 up to 50% of all production operating expenses (Tompkins and White, 1984).

T
According to the same authors, one can decrease these material handling expenses by as much as 10 to

IP
30% by improving facility layouts. In studies regarding man-machine system designs, the optimization

CR
is focused on minimizing psycho-physical costs of human work. Efficiency criteria in this regard may

involve a number of performed body parts movements, distance covered by a person, visual work

US
intensity measured by means of fixations and saccades, etc. (e.g., Sargent et al., 1997).
AN
Generally, the optimization task may be formulated in such a way:
M

n 1

  D  Lij 
n
Minimize: p (i ) p ( j ) (1)
i 1 j i 1
ED

where Lij is the relationship strength between two objects and Dp(i)p(j) is the distance between item

location places p(i)p(j).

PT

Despite the simplicity of the problem description, finding the best solution is not a trivial task. Sahni
CE

and Gonzalez (1976) proved that the FLP is an NP-hard problem and, thus, it is impossible to obtain

an optimal layout in polynomial time. As a consequence, a great number of various sub-optimal

AC

methods have been developed through years. In practice, these techniques provide satisfactory

solutions in acceptable time. Various mathematical formulations of this problem are surveyed by Loiola

et al. (2007). Extensive reviews of different methods of finding solutions were presented in numerous

works, for example, Kusiak and Heragu (1987), Liggett (2000), Drira et al. (2007) or lately Kundu and

Dan (2012) and Moslemipour et al. (2012).

2
 ACCEPTED MANUSCRIPT

Many approaches to solving FLP may be generally divided into two classes: constructive and

improvement-based. In the former group, consecutive objects are inserted into the layout one by one

(e.g., ALDEP proposed by Seehof et al., 1966 or CORELAP created by Parsaei and Galbiati, 1987)

whereas in the latter case, the initial arrangement of all items is repeatedly improved by exchanging their

locations (e.g., CRAFT). Hybrid methods involve both techniques. It was initially proved by Liggett

(1981) that using results from constructive algorithms as an initial layout for an improvement procedure

yields better results than starting from a random layout. For more discussion in this regard see Liggett

T
IP
(2000). Metaheuristics constitute another category of hybrid optimization methods. Many proposals in

this respect appeared in the last three decades. Among them there are: tabu search (Taillard, 1991;

CR
Skorin-Kapov, 1994), simulated annealing (Heragu and Alfa, 1992; Jajodia et al., 1992; Matai, 2015;

US
Palubeckis, 2015), particle swarm approach (Önüt et al., 2008; Ohmori et al., 2010; Samarghandi et al.,

2010), artificial immune system algorithms (Kumar et al., 2008; Haktanirlar Ulutas and Kulturel-Konak,
AN
2012), genetic algorithms (Sadrzadeh, 2012; Aiello et al. 2012, 2013; Gonçalves & Resende, 2015; Lim

et al., 2017; Paes et al., 2017; Palomo-Romero et al., 2017), or ant colony optimizations (Komarudin
M

and Wong, 2010; Guan & Lin, 2016).

ED

The described above FLP classes of models have a quantitative nature which can be regarded as an

advantage from the computational point of view. However, preparing a good model of a real FLP in
PT

such a perspective is not easy. The difficulties arise, generally, from two reasons: the necessity of

including qualitative parameters and imprecision of quantitative data. These problems inclined
CE

researchers to seek solutions that use experts‘ knowledge in different stages of making decisions about

objects‘ locations. One of the first approaches in this spirit was the proposal of Muther (1961) in which
AC

the inter-relations are determined by experts who subjectively assess the importance of objects‘

closeness. The importance and benefits of the collaboration between quantitative algorithms and

experts‘ feedback were shown, for instance, by Babbar-Sebens & Minsker (2012) in their interactive

genetic algorithm. Recently, a different approach of applying expert‘s knowledge and preferences while

solving facility layout problems was proposed by García-Hernández et al. (2013). They put forward a

specific version of the genetic algorithm, in which the designer was repetitively involved in subjectively
3
 ACCEPTED MANUSCRIPT

assessing representative solutions derived by clustering from a given population. This interactive

technique was further extended in the work of García-Hernández et al. (2015a) where a multi-objective

version of their genetic algorithm was put forward. Improvements of the original idea from the paper

García-Hernández et al. (2013) were also presented in the study of García-Hernández et al. (2015b).

The authors additionally included two niching techniques into the approach to prevent presenting

similar solutions in the same iteration of their interactive genetic algorithm. A work of Zhao et al.

(2014) can also be included in this research direction. The particle swarm optimization-based immune

T
IP
algorithm proposed in this paper allows for including in the initial population possible solutions

(individuals) supplied by an expert. However, in this case, the expert‘s knowledge and experience are

CR
not used in further steps of the procedure.

US
Models of FLP for inexact data, intensively developed in recent decades, often involve fuzzy sets theory

components. Evans et al. (1987), for instance, developed a fuzzy heuristic of constructive nature based
AN
only on qualitative data. The procedure aimed at determining the order of departments while the

placement was conducted manually. Dweiri and Meier (1996), in turn, used the fuzzy decision-making
M

system to produce activity relationship charts that were later applied to develop layouts of departments
ED

in a plant. Cheng et al. (1995) proposed a genetic algorithm where the material flow was represented by

trapezoidal fuzzy members. Another version of a genetic algorithm involving fuzzy logic was developed
PT

by Karray et al. (2000) to deal with temporary facilities on a construction site. The work of Aiello and

Enea (2001) focuses on solving facility layout problems in the uncertainty of the market demand
CE

condition which was modeled by triangular fuzzy numbers. They took into account departments‘

production capacities. A comprehensive, fuzzy based, decision support system was proposed by Deb
AC

and Bhattacharyya (2005). In their multi-factor fuzzy system, they included material flow, as well as

supervision, information, and environmental links. Bashiri & Hosseininezhad (2008) examined the

optimal location of new facilities with respect to a set of existing ones by mixing the fuzzy analytic

hierarchy process approach with fuzzy linear programming.

Some studies proposed hybrid approaches combining both qualitative data and linguistic components

provided by experts: e.g., Sangwan (2010) solved three plant layout problems by means of his multi-
4
 ACCEPTED MANUSCRIPT

criteria approach, Mohamadghasemi & Hadi-Vencheh (2012) integrated fuzzy judgments and nonlinear

programming to assess facility layout structures, and recently Altuntas et al. (2014) applied their

proposal to solve a real problem in a machinery industry company. Another multi-criteria approach

including a mix of various techniques for solving problems with non-crisp, operational, qualitative, and

dependent indicators was proposed by Azadeh et al. (2016). They combined fuzzy simulations and

fuzzy data envelopment analysis with Fuzzy Analytic Hierarchy Process which was used to take

advantage of expert‘s knowledge of imprecise data.

T
IP
A specific application of the multi-valued logic in conjunction with fuzzy sets components was

CR
proposed in the works of Grobelny (1987a, 1987b, 1988). The framework called Linguistic pattern method

(LP) allows utilizing logical expressions as criteria in solving facility layout problems. Imprecise data can

US
be formulated as expressions similar to a natural language. Raoot and Rakshit (1991, 1993) developed

multiple criteria constructive algorithm based on subjective experts‘ opinions by taking advantage of
AN
this framework. Recently, in the work of Grobelny (2016), it was shown how to apply the concept of

linguistic patterns to determining the objects‘ hierarchy based on a linguistic assessment of pairwise
M

comparisons.
ED

This work presents a novel implementation of the simulated annealing involving the idea of linguistic

patterns. The new algorithm is applied to solve facility layout problems on regular grids where objects
PT

may be modeled by a single cell or a group of cells. As far as we are aware none of the previous studies

used linguistic patterns representing either experts‘ subjective knowledge or imprecise data as a goal
CE

function in the simulated annealing algorithm. In experiments involving real-life and artificial examples,
AC

we show that our method yields not only logical, reasonable, and satisfactory solutions but also

provides results that are comparable with layouts produced by a classic procedure for crisp and exact

input.

Next sections explain the concept of linguistic patterns on some simple examples and describe in detail

our simulated annealing modification. Further, we show the results of simulation experiments regarding

the proposed approach. The goal of the first set of experiments was meant to identify parameters‘

5
 ACCEPTED MANUSCRIPT

values that maximize the effectiveness of the algorithm. The obtained outcomes are then used in the

follow-up, set of experiments that examine how our proposal performs in various problem types

involving approximate experts‘ assessments of objects‘ relationships. These investigations include

objects both with uniform and variable sizes. The results of our LP approach are directly compared

with an analogous crisp version of the simulated annealing procedure. The discussion is specifically

focused on the conceptual differences between these two versions.

T
2. Linguistic pattern as a criterion in facility layout problems

IP
There are many sources of imprecision in facility layout problems. From one hand side, while searching

CR
for an optimal layout in practical tasks it is necessary to use qualitative criteria or variables that are

difficult to measure. For instance, Garcia et al. (2013) included into this category: the preferred

US
location, a division of free space or environmental conditions. In turn, Altuntas et al. (2014) mention:

personnel, equipment, and information flow or supervision. On the other hand, during the layout
AN
design, sometimes even quantitative data can be determined only approximately. They may be specified

according to the knowledge, intuition, and experience of experts. It applies, for example, to degrees of
M

relationships between objects being placed. Muther (1961) proposed a RelCharts technique that

addresses this issue. In this approach, experts assess the necessity of locating objects next to each other
ED

on a scale based on six linguistic expressions. For some facility layout criteria, the assessment of a
PT

physical variable may be flexible which means that a specific value may have a different meaning for

various criteria. For instance, a given distance between two objects has a different meaning for
CE

transport costs and for the supervision process quality. In contrast to the role of an expert in Garcia et

al. (2013, 2015a, 2015b), the LP approach is generally directed to the use of expert‘s knowledge for
AC

modeling imprecise criteria and data before performing the optimization algorithms.

Generally, the linguistic pattern can be specified as a logical phrase including linguistic variables. The

characteristic feature of these variables is that one can assign verbal values to them like in a natural

language. The adjective logical implies that it is possible to assess the truth or a degree of truth for a

given sentence like in the expression IF A THEN B. In the domain of facility layout problems, the

basic optimization task defined in (1) can also be formulated by natural language expressions. For
6
 ACCEPTED MANUSCRIPT

instance, one may say that in a good arrangement, for any pair of objects, the following relation is true:

if the positive relationship between items is strong, then they are located close to each other. By putting this rule in

a more formal form we get the following linguistic pattern:

IF Link_between_objects(i, j) is STRONG THEN Distance_between_objects(i, j) is SMALL (2)

Since formula (2) presents an implication, we can evaluate the level of its truth. For example, if strongly

T
connected items are placed distant from each other, the whole sentence is false.

IP
In the current paper, Truth_of(l) and Truth_of(r) refer to truth values of left and right sides of linguistic

CR
pattern (2), that is Link_between_objects(i, j) is STRONG and Distance_between_objects(i, j) is SMALL,

respectively. The most general approach to determining the left and right-hand side truths was

US
proposed by Zadeh (1978a, 1978b). He introduced the consistency term specified as a relative truth level

for the expression ―E is A‖ with respect to ―E is B.‖ where expressions are represented by fuzzy
AN
numbers or sets. Such a definition of consistency can be computed as a possibility value according to the

following formula:
M
ED

CONSISTENCY(E is B, E is A) = POSSIBILITY(E is B | E is A) = maximumx(minimum(B(x), A(x)) (3)

PT

where X is the universe of discourse, B(x) and A(x) are the membership functions of B and A,
CE

respectively, and | denotes with respect to. Thus, for the left side of pattern (2), we may write the

following.
AC

Truth_of(l) = POSSIBILITY (Link_between_objects(a, b) is STRONG | Link_between_objects(a, b) is A) (4)

Truth_of(l) = maximumx(minimum(STRONG(x), A(x)) (5)

7
 ACCEPTED MANUSCRIPT

The right side of pattern (2) can be expressed analogically. A graphical illustration of obtaining the

consistency value for A = MEDIUM is given in Fig. 1.

Membership
(truth) value

1 MEDIUM(x) STRONG(x)

Truth_of(l) =
.5 POSSIBILITY(
Link_between_objects(a, b) is STRONG |

T
Link_between_objects(a, b) is MEDIUM)

IP
0 5 10 (x)

CR
Fig. 1. Graphical illustration of exemplary definitions of linguistic variables and the calculation of
Truth_of(l) according to equation (5).

US
As it was shown in a series of works published by Grobelny (1987a, 1987b, 1988, and 2016), one can

assess the truth of pattern (2) by means of the Łukasiewicz formula which takes the following form:
AN

Truth_of(pattern (2)) = minimum{1, 1 - Truth_of(l) + Truth_of(r)} (6)

M

It seems that, in practical applications, the easiest way of obtaining truth assessments for expressions on
ED

the right side of formula (6) is to directly ask experts. The quality of the given layout configuration may

be then measured by computing values from formula (6) for all pairs of objects and calculating their
PT

arithmetic average. This is treated as a goal function in our version of simulated annealing.
CE

∑ ∑ ( )
Maximize pattern mean truth: (7)
AC

Since for a specific arrangement it is often unfeasible to obtain the full mean truth which is equal one, it

is a good idea to relate the current solution mean truth to the maximum possible truth for a given

facility layout problem. Grobelny (1987b) provided a theorem showing how such an upper bound

should be determined.

8
 ACCEPTED MANUSCRIPT

The following, very simple, illustrative example shows how to use pattern (2) as an optimization

criterion in facility layouts where relationships are approximated and distances between objects have

linear meaning. Let us assume that there are three machines arranged in a row. It is also known how

many transport operations between these machines need to take place during a specific period (e.g., one

hour). Exemplary data along with two possible layout solutions are provided in Fig. 2.

L U V W
U × 9 6

T
Layout 1: U W V
V × 1

IP
W × Layout 2: W V U
Fig. 2. An exemplary, simple facility layout problem where relationships between objects are based on the

CR
number of transport operations between each pair of facilities in the sequence ij or ji, and two possible layouts
on a modular grid (1×3).
An expert directly assessed truth values of the linguistic expression STRONG depending on the

US
number of possible transport operations as well as the expression SMALL for the distance which, in

this specific layout context, is defined as the number of machines between the given pair. The data are
AN

illustrated in Figs. 3 and 4 respectively.

M

Membership
(truth) value
ED

STRONG(y)
1
.8
.6
PT

.4
.2
CE

V.Small Small Medium Big V.Big Degree of link (y)

{1, 2} {3, 4} {5, 6} {7, 8} {9, 10} Crisp value of link
(e.g., number of transport operations)
AC

Fig. 3. The pattern STRONG for a degree of relationships (e.g., number of transport operations) between two
objects in a form of fuzzy set singletons that were intuitively defined by directly assigning the truth values.

9
 ACCEPTED MANUSCRIPT

Membership
(truth) value

1
SMALL(y)

.5

0 1 2 Distance (y)
No of objects between pair of objects

T
Fig. 4. The pattern SMALL for a distance between machines in a form of fuzzy set singletons that were defined
by directly assigning the truth values.

IP
Given the problem data specified in Fig. 2 and linguistic pattern definitions shown in Figs. 3 and 4, we

CR
can compute truth values for all pairs of objects in both layouts using Łukasiewicz formula (6).

US
Assuming unitary distances between adjacent machines we get:
AN
Layout 1: U W V Layout 2: W V U
Truth_of {U, V} minimum{1, 1 – 1 + .5} = .5 minimum{1, 1 – 1 + 1} = 1
Truth_of {U, W} minimum{1, 1 – .6 + 1} = 1 minimum{1, 1 – .6 + .5} = .9
M

Truth_of {V, W} minimum{1, 1 – .2 + 1} = 1 minimum{1, 1 – .2 + 1} = 1

ED

The average truth for Layout 1 amounts to 2.5/3 = .83 while for Layout 2: 2.9/3 = .97, thus, the second

arrangement fulfills the pattern (2) better than the first configuration which means that Layout 2 is
PT

better in terms of the mean truth value from equation (7).

The linguistic pattern methodology allows also for taking into account most of the earlier mentioned
CE

cases in modeling facility layout tasks. It results from the fact that usually imprecise and indistinct
AC

criteria may be described by logical sentences. Furthermore, the majority of variables used in these

criteria can be represented by degrees of intensity expressed linguistically. One may approximate

physical values as well as use diverse meanings of linguistic expressions in various design contexts by

applying fuzzy sets for these expressions. Therefore, it is possible to apply optimization based on the

greatest mean truth value of specified criteria defined by linguistic patterns. Analyzing, for instance, the

qualitative criteria postulated by Altuntas et al. (2014), it is easy to notice that they may be depicted by

10
 ACCEPTED MANUSCRIPT

implication expressions (IF A THEN B). If one assumed that information flow relates to the intensity

of necessary personal contacts between machines‘ operators i and j, then the layout quality pattern

could be described as follows:

IF Degree_of_contacts(between operators of objecti and objectj) is HIGH

THEN (objecti and objectj are located NEAR)

In turn, if the environmental condition criterion denotes inconvenience or danger resulting from the

T
adjacency of objects i and j, then a possible pattern may take the following form:

IP
IF Degree_of_hazard(adjacency of objecti and objectj) is HIGH

CR
THEN (objecti and objectj are NOT ADJACENT)

US
To compute truth values according to formula (6), it is necessary to define appropriate patterns for left

and right sides of implications as well as values of levels for variables: Degree_of_contacts(between
AN
operators of objecti and objectj) and Degree_of_hazard(objecti and objectj adjacency), respectively, for all

ij objects‘ pairs. The same applies to expressions NEAR and NOT_ADJACIENT.

M

Undoubtedly, all the mentioned parameters are of an imprecise nature and their definitions can be
ED

provided by experts. It is clear that models obtained in such a way would include components of

subjective knowledge and experience of the decision maker. This, in turn, may have an impact on the
PT

precision and representation of linguistic pattern items. Figs. 5 and 6 show possible definitions of

patterns for left sides of described linguistic expressions while Figs. 7 and 8 present examples for their
CE

right side counterparts. From one hand side, they show the flexibility of defining the model in linguistic

pattern categories and, on the other, present how to take advantage of the more or less precise experts‘
AC

knowledge.

11
 ACCEPTED MANUSCRIPT

Membership
(truth) value
HIGH Degree_of_hazard(y)
(or Number_of_contacts(y))
1
.5

.2

Small Medium Big Expert‘s assessment (y)

Fig. 5. An exemplary truth value definition of HIGH Degree_of_hazard(y) (or Number_of_contacts(y)) in a form of fuzzy set

T
singletons defined by directly assigning truth values.
Membership

IP
(truth) value
HIGH Degree_of_hazard(y)
(or Number_of_contacts(y))

CR
1

.5

.2 US
AN
60 120 noise dB(A) (y)
30 60 contacts (minutes)
Fig. 6. An exemplary truth value definition of HIGH Degree_of_hazard(y) (or Number_of_contacts(y)) in a form of a continuous
fuzzy set function specified by experts.
M

Membership
ED

(truth) value
NEAR(y) NOT_ADJACENT(y)
PT

1
.5
CE

.2

0 1 2 Number of objects
AC

between pair (y)

Fig. 7. Exemplary truth value definitions of NEAR and NOT_ADJACENT objects‘ locations in a form of fuzzy set
singletons defined by directly assigning truth values.

12
 ACCEPTED MANUSCRIPT

Membership
(truth) value

NEAR(y) NOT_ADJACENT(y)
1

.5
.2

Distance between
2 4 6 8 10 centroids in meters (y)

Fig. 8. Exemplary truth value definitions of NEAR and NOT_ADJACENT in a form of a continuous fuzzy set function

T
specified by experts.

IP
CR
3. Linguistic pattern based simulated annealing

US
The classic simulated annealing initially proposed by Kirkpatrick et al. (1983) is inspired by the way, the

metal particles behave during the annealing process. This movement is simulated by changing items‘
AN
locations. Generally, the change in objects‘ locations takes place if it improves the goal function value.

However, the algorithm allows also for making an unfavorable change with a given probability (P).
M

  
 
 T 
P(accept unfavorable change) = e , (8)
ED

In practice, apart from defining this probability, other parameters need to be specified, that is, the initial

temperature (Ti), epoch length, cooling schedule, and termination criterion.

PT

A number of recommendations concerned with setting these values are comprehensively reviewed by
CE

Singh and Sharma (2008). We have taken advantage of these suggestions while implementing our

approach. Kirkpatrick et al. (1983) proposed to set the probability of accepting worse solutions at
AC

approximately .8 while starting the procedure, which should help in avoiding getting stuck in local

optima. In our algorithm, before the simulated annealing procedure begins, first, N2 (N is the number

of all objects) random changes of objects pairs are made and the mean value of ∆s is computed. Based

on this and any predefined probability value, we calculate the initial temperature Ti individually for a

specific facility layout problem. A comparable procedure was also applied by Singh and Sharma (2008).

In the current implementation the Epoch Length, which specifies the number of objects‘ location

13
 ACCEPTED MANUSCRIPT

changes in each temperature, can be set either arbitrarily or as a multiplier of the number of objects in

the current problem (k × N). The latter solution was used in Wilhelm and Ward (1987) or Heragu and

Alfa (1992). Regarding the cooling scheme, we applied the formula used by Heragu and Alfa (1992),

that is: Tj = r × Tj-1, where r is usually set at .9 and can be freely specified in our application. The final

temperature is obtained after i steps (usually 100) according to the following formula: Tf = .9(i - 1)×Ti.

As compared to the conventional simulated annealing algorithm, our main extension is the possibility

T
of entering data and calculating the objective function according to the idea of linguistic patterns

IP
(Grobelny 1987a, 1987b, 1988, 2016) described in the previous section. More specifically, one can

CR
define appropriate linguistic variables in fuzzy set terms as it was illustrated in examples from section 2.

Our implementation of the LP based simulated annealing algorithm allowed for conducting

US
experiments focused on properties of the proposed approach in various contexts and definitions of the

facility layout tasks.

AN
Drira (2007) proposed a sophisticated system for FLP classification based on three groups of factors,

namely the workshop characteristics, problem formulation, and the approach to finding the solution.
M

The workshop characteristics usually include production and transport systems (Kusiak and Heragu,
ED

1987). The problem definition involves objects and design data representations while the solution

approach is specified by various types of optimization algorithms.

PT

In the developed application, a discrete representation of objects placed on the modular grid was
CE

assumed. Thus, the constraints of our approach concern objects‘ locations, their sizes, and available

space modeled as a regular and rectangular grid with fixed dimensions H(No of Cells × W(No of Cells), where grid
AC

height (H) and width (W) are expressed as a number of cells. Naturally, the number of available cells in

a grid must be equal or bigger than the sum of cells representing objects (O) or objects‘ components.

This is mathematically expressed by formula (9). Additionally, objects should not overlap.

∑ (9)

14
 ACCEPTED MANUSCRIPT

This type of representation is the most suitable for modeling cellular layouts with an open field

transportation system. An optimization of the mutual arrangement of cells having identical dimensions

is the simplest case that can be analyzed under such assumptions. Jajodia et al. (1992) showed that it is

possible to consider layouts consisting of objects with various sizes by appropriately defining cells‘

relationships. In our proposed application, such a functionality can also be achieved by constructing

complex objects from elementary cells. The software calculates both Manhattan and Euclidean

distances. They correspond to a basic transport system that is, conveyors or Automatically Guided

T
IP
Vehicle (AGV) and gantry robots in the latter case (Kusiak and Heragu, 1987). The computer program

also enables modeling other transportation systems by means of inserting artificial objects and blocking

CR
their positions. The quality criterion in the proposed algorithm has the form of implication (IF A

THEN B).
US
Data in the relationship matrix can have either the form of physical quantities or linguistic expressions
AN
whereas left and right side pattern definitions are constructed by means of providing trapezoidal

membership functions. Fig. 9 shows a part of the user interface where exemplary membership
M

functions for the left and right side of pattern (2) are specified linearly, that is, for the expressions
ED

Link_between_objects(i, j) is STRONG and Distance_between_objects(i, j) is SMALL respectively.

PT
CE
AC

Fig. 9. Defining fuzzy number shapes used for computing the fuzzy goal function values.

15
 ACCEPTED MANUSCRIPT

The mean truth value of the pattern (2), calculated for all pairs of linked objects, constitutes the

criterion (formula (7)) being analyzed at every step of the proposed simulated annealing procedure.

Using the simulated annealing algorithm for analyzing problems where objects have unequal sizes

requires special numerical solutions. Jajodia et al. (1992) suggested that object building blocks should be

connected with each other with artificial links of the magnitude being 1.5 - 2 times bigger than the

T
maximum link value from the initial matrix. Such a solution should keep objects‘ components together

IP
while applying the algorithm.

CR
We propose a slightly different approach in our application. We treat one of the building blocks of

a given object (facility) as the input/output point. Relationships between these components reflect the

US
transportation system inside this object. Such a model allows for analyzing patterns‘ truths both for

relationships between objects and inside them. Examples are given in Figs. 10 and 11.
AN
M
ED
PT
CE

Fig. 10. A possible way of relating two objects Fig. 11. A possible way of relating two objects
consisting of 5 and 4 building blocks (grid cells) with consisting of 5 and 4 building blocks (grid cells) with
AC

links expressing the truth value of the pattern (2) left links expressing the truth value of the pattern (2) left
side for the open field type of transportation model. side for the loop type of transportation model.

The prepared software provides truth values for both types of transportation systems. It is also possible

to perform a local optimization of layouts inside objects both in the classic simulated annealing and in

its version based on linguistic expressions. Objects represented by a number of smaller items may

disintegrate during the optimization process. Therefore, an additional algorithm is being used to check

16
 ACCEPTED MANUSCRIPT

their integrity. The software registers the best solutions from a series of automatically repeated

simulations along with measures of objects‘ disintegration degrees.

For the sake of making comparisons between the traditional and the LP approach available, the

software computes classic goal functions for solutions obtained by means of the algorithm using

linguistic patterns and vice versa.

4. Simulation experiments

T
The objectives of the experiments reported here were twofold. Because it is possible to determine the

IP
upper bound for the pattern (2) mean truth value according to the maximal mean truth theorem

CR
(Grobelny 1987b), one can assess the quality of the given solution in relation to this upper bound mean

truth values. Thus, in the first research perspective is focused on examining the sensitivity of solutions

US
to parameters controlling the simulated annealing process. Secondly, it is essential to investigate

relationships between variables and functions in the presented novel approach and a classic, crisp
AN
version of the simulated annealing algorithm. Both sets of experiments investigate the behavior of our

LP implementation with the goal function in a form of equation (7). These experiments and obtained
M

simulation results are presented and described in the following subsections.

ED

4.1. Properties of the proposed algorithm

The first experiment was designed to investigate the influence of algorithm parameters on the
PT

effectiveness of finding solutions measured by the mean truth value (7), that is, the average truth value
CE

of pattern (2) computed for all pairs of objects, where Truth_of(l) > 0. Examined parameters included

the probability of accepting worse solutions (P), the multiplier used for determining the Epoch Length (k)
AC

as well as the r parameter which is responsible for the cooling speed scheme. For each of these factors,

two levels were specified: small and big.

A specific version of the typical facility layout problem was prepared to compare LP and crisp types of

simulated annealing algorithms. The idea is as follows. Let us assume that there is a system of machines

or tools with a given relationship matrix that includes data about the intensity of interactions for each

objects‘ pair (Lij) in a form of physical values (e.g., transportation operations). Such a problem

17
 ACCEPTED MANUSCRIPT

formulation may be described as objective. On the other hand, an expert, based on his experience and

knowledge of this system, is capable of determining the relationship matrix by means of linguistically

expressed levels of links on an ordinal scale with a granularity equal five (Very Small, Small, Medium, Big,

Very Big) or three (Small, Medium, Big). The expert is rational in such a sense that his linguistic

assessments cover appropriate succeeding ranges of physical values divided into five or three equal

parts. When Lij values are natural numbers from one to ten, then the more precise expert distinguishing

five levels will assign the Very Small level to values one and two, the Small to two and three, etc.

T
IP
For simplicity, in all experiments, the left side of pattern (2) is operationalized by a linear assignment of

CR
truth levels Truth_of(l) to expressions obtained as in Fig. 3. This means that in simulations for

granularity five it was assumed that Truth_of(Very Small) = .2, Truth_of(Small) = .4 etc. whereas for

US
granularity three: Truth_of(Small) = .1, Truth_of(Medium) = .5, and Truth_of(Big) = 1. The truth value of

the right side of pattern (2) is generally determined in accordance with the idea presented in Fig. 4, but
AN
here we used a continuous linear function specifying the maximal possible Manhattan distance on a

regular grid applied in a given problem type.

M

Likewise other metaheuristics, the simulated annealing procedure belongs to the improvement type of
ED

algorithms. This category of techniques performs better than construction methods in difficult facility

layout problems with a small value of flow dominance and a big density of links (Kusiak & Heragu
PT

1987). Therefore, for our experiments, we chose well-known Nugent 30 and Nugent 16 problems

structures from the work of (Nugent et al., 1968), where relationships are expressed by natural numbers
CE

from one to ten, their density amounts to about 70%, and the flow dominance - approximately 1.2.
AC

Additionally, we generated two problems with a comparable number of objects to the selected

Nugent‘s examples, that is, 4×4 and 6×6 with randomly assigned link values from the range of 1 – 10, a

similar flow dominance value but with a lower link matrix density (40%). Linguistic patterns for objects‘

relationships, in all examined examples, were specified for granularity five and three according to

rational expert rules described above.

18
 ACCEPTED MANUSCRIPT

Formally, the full factorial, within subjects design produced 16 different experimental conditions,

namely P{.1, .8} × k{10, 20} × r{.90, .99} × Gr{3, 5}. For each of the four problem types {Nugent 16,

Nugent 30, Random 16, Random 36}, and experimental condition the algorithm was repeated 100 times.

Table 1 and Figs. 12-15 present the main results obtained by the classic factorial analysis of variance.

The lack of significant difference between two employed levels of probability of accepting worse

solutions (P) is rather surprising (Fig. 14). Such a result has not been previously reported and, to

T
authors‘ knowledge, this factor has not been yet systematically analyzed in the literature on the

IP
simulated annealing algorithm and its modifications. A distinct and strong (a high value of partial η2)

CR
impact of the r value, which controls the cooling process has not been earlier demonstrated. Usually,

the parameter is set at .9 without any broader discussion while the result illustrated in Fig. 13 suggests

US
that the level of .99 provides meaningfully better solutions in our implementation of simulated

annealing. Obtaining considerably higher mean values of the corrected truth for the smaller granularity
AN
(3) is rather intuitive since the algorithm searches, in this case, a smaller problem space than in the

bigger granularity condition (Fig. 12). Similarly, better results observed for the higher Epoch Length
M

Multiplier (k = 20) were also expected as the bigger value of Epoch Length signifies that the algorithm
ED

performs more pairwise changes for every temperature and, thus, the probability of finding a better

layout increases (Fig. 15). It is worth noting that the investigated parameters do not interact with each
PT

other (Table 1).

Table 1. Four-way ANOVA results for the first experiment. The effects of Granularity (Gr), Cooling Scheme (r),
CE

Probability (P) of accepting worse solutions, and Epoch Length Multiplier (k) on the mean corrected truth value for
pattern (2) in four types of facility layout problems {Nugent 16, Nugent 30, Random 16, Random 36}.
Effect SS df MSS F p η2
AC

Granularity (Gr) .064 1 .064 589 <.0001 .11

Cooling Scheme (r) .0046 1 .0046 43 <.0001 .0089
Probability (P) 0 1 .000 0 1
Epoch Length Multiplier (k) .00059 1 .00059 5.4 .020 .0011
Gr × r .00015 1 .0001481 1.4 .24
Gr × P .0000028 1 .0000028 .026 .87
r×P .0000022 1 .0000022 .020 .89
Gr × k .0000027 1 .0000027 .025 .87
r×k .000068 1 .0000676 .62 .43
P×k .00000094 1 .00000094 .0087 .93
Gr × r × P .0000012 1 .0000012 .011 .92
Gr × r × k .00000063 1 .00000063 .0058 .94
19
 ACCEPTED MANUSCRIPT

Gr × P × k .0000065 1 .0000065 .060 .81

r×P×k .00000035 1 .00000035 .0033 .95
Gr × r × P × k .00000052 1 .00000052 .0048 .94
Error .52 4784 .00011

0.990 0.986

Mean corrected truth value

Mean corrected truth value

0.988
0.985
0.986
0.984
0.984

T
IP
0.982 0.983

0.980

CR
0.982
3 5 0.90 0.99
Granularity Cooling Scheme (r)

F(1, 4784) = 589, p < .0001, η = .11.

2
US
Fig. 12. The influence of Granularity on mean corrected Fig. 13. The influence of Cooling Scheme r parameter on
truth values for pattern (2). mean corrected truth values for pattern (2).
F(1, 4784) = 43, p < .0001, η2 = .0089.
AN
Vertical bars denote .95 confidence intervals. Vertical bars denote .95 confidence intervals.

0.9850 0.986
Mean corrected truth value

Mean corrected truth value

M

0.9845 0.985
ED

0.9840 0.984
PT

0.9835 0.983
CE

0.1 0.8 10×N 20×N

Probability Epoch Length

Fig. 14. The influence of Probability of accepting worse Fig. 15. The influence of Epoch Length Multiplier (k) on
AC

solutions on mean corrected truth values for pattern (2). mean corrected truth values for pattern (2).
F(1, 4784) = 0, p = 1. F(1, 4784) = 5.4, p = .020; η2 = .0011.
Vertical bars denote .95 confidence intervals. Vertical bars denote .95 confidence intervals.

4.2. Comparison of LP and crisp simulated annealing

The main goal of the second set of experiments was to verify the quality of algorithmic solutions based

on the experts‘ knowledge expressed by means of linguistic patterns and linguistic expressions. In the

proposed approach we assumed that there exists an objective system of assessing costs expressed by
20
 ACCEPTED MANUSCRIPT

function (1). The expert doesn‘t exactly know that system but has experience and approximate

knowledge in this regard that can be formulated as logical patterns similar to a natural language. He can

specify the intensity of transportation with a given precision in linguistic terms. Such an assessment can

be made on an ordinal scale.

The confrontation of solutions obtained for our linguistic input and results provided for classic,

objective data can be treated as a practical verification of the quality and usability of the proposed

T
approach. Since broadly defined transportation costs may be used as a strict indicator of the layout

IP
quality, our experiments‘ relationship matrices include transportation operations‘ intensities. Function

CR
(1) computes the cost of these operations whereas formula (7) is its linguistic counterpart.

The first experiment, in this second research perspective, investigates examples representing

US
intracellular transportation tasks for cells of identical shapes and dimensions and the open field

transportation type, where the distance is computed according to the Manhattan measure.
AN
The second experiment deals with more sophisticated models including objects with unequal areas.
M

Additionally, we try to optimize transportation costs not only between objects but also inside them.

Here, we also assumed the open field type of transportation, but distances were calculated not only by
ED

the Manhattan measure but also according to the Euclidean formula which simulates the gantry robot

behavior.
PT

4.2.1. Objects with uniform sizes

CE

This experiment was designed to investigate the behavior of the proposed method in comparison with

the classical optimization approach which minimizes function (1) and assumes that objects have equal
AC

dimensions. For this purpose, we used again the same well-known examples Nugent 16 and Nugent 30

along with random matrices 4×4 and 6×6 as in the first set of experiments. In addition, we used an

8×8 matrix with randomly generated links from the 1-10 range, similar density (40%) and flow

dominance (1.4). For these matrices, the best solutions were found by means of classic simulated

annealing with a standard goal function defined according to formula (1) and distances measured by

a Manhattan metric. Taking advantage of the first set of experiments‘ results, the parameters were set at

21
 ACCEPTED MANUSCRIPT

P = 0.8, r = .99 and k = 20. For each experimental condition, the procedure was repeated 100 times.

Next, the link data were transformed to expert‘s opinions like in the first set of experiments by means

of three and five levels of linguistic assessments of links. Again, for each type of examined problems,

the best solutions in terms of pattern (2) mean truth value were searched for. For these best layouts, we

computed values of function (1) for original link data. The obtained results were compared with the

best classic solutions (optimal for Nugent 16 and Nugent 30) and put together in Table 2.

T
These outcomes primarily show that the layout solutions, obtained by the algorithm with linguistic

IP
patterns, not only exhibit high mean truth levels but are also good in terms of objective data and goal

CR
functions computed according to formula (1). This is especially evident for more precise experts

assessing with granularity five, where they differ by merely two to six percent from the best solutions

US
produced by the classic simulated annealing algorithm. What is quite interesting, these differences seem

not to depend on the problem size, which apparently is the case for mean truth levels.
AN
Table 2. Results for examples with uniform objects‘ sizes. Upper bounds were computed according to the
theorem about maximal mean truth (Grobelny, 1987b). The corrected values are calculated as a difference
between mean Truth_of(pattern (2)) and the upper bound.
M

Mean Truth Values

Mean Truth Values Goal Function LP
for Best Crisp
Problem for LP approach Values to Crisp
Granularity Solutions
type Deviation
ED

Upper Best
Pattern (2) Corrected LP [%] Pattern (2) Corrected
bound Crisp
4×4 3 .934 .955 .979 424 382 11.0 .917 .962
5 .957 .973 .984 390 382 2.1 .921 .948
PT

6×6 3 .935 .964 .971 4130 3851 7.2 .925 .961

5 .931 .982 .949 4037 3851 4.8 .922 .940
CE

8×8 3 .920 .964 .956 19870 18835 5.5 .908 .944

5 .919 .973 .947 19400 18835 3.0 .911 .939
Nug_16 3 .973 .978 .995 681 620 9.8 .967 .989
5 .962 .978 .984 660 620 6.5 .958 .980
AC

Nug_30 3 .981 .989 .992 3196 3062 4.4 .977 .988

5 .974 .989 .985 3123 3062 2.0 .969 .980

The observed discrepancies between compared simulated annealing approaches do not result solely

from the approximations concerned with transforming real data to experts‘ opinions having predefined

granularity levels. A comparison of the results for classic and fuzzy implementations presented in

columns 3 and 9 of Table 2, indicates that both optimization criteria do not work identically, despite

22
 ACCEPTED MANUSCRIPT

reflecting a similar idea. The linguistic pattern version, along with its parameters and definitions of

expressions, is not a simple generalization of the classic approach. It may be noticed that optimal

solutions in a classic sense are not the best in terms of the mean truth and conversely. The main reason

probably lies in a nonlinear character of the assessment model expressed in a form of the truth level of

pattern (2). One can easily note that the pattern (2) truth level criterion computed according to the

Łukasiewicz (6) formula is sensitive to changes in the mutual arrangement of a given pair only if

Truth_of(l) > Truth_of(r). This means, for instance, that the relative location of objects (i, j) with a link

T
IP
level equal Medium, for which Truth_of(l) = .5 is evaluated exactly the same for distances from the range

of 1% to 50% of the maximum possible distance. In the current paper experiments, we used regular

CR
grid model with unitary distances between cells. For the Manhattan measure used in our experiments,

US
the maximal distance can be determined by the formula: (p + q) – 2, where p and q specify the grid

dimensions. In the example with a 4 × 4 grid, the maximal distance equals six, so the full truth of
AN
pattern (2) will occur if the distance between objects would be one, two, or even three units because

Truth_of(3) = .5. The calculations assume a linear way of evaluating the truth of pattern
M

Distance_between_objects(i, j) is SMALL.
ED

Obviously, the biggest sensitivity to locations‘ changes will occur for pairs with the maximum truth of

the left-hand side of pattern (2) (Truth_of(l) = 1). In such an event, the distance one is optimal and any
PT

increase of this value would decrease the truth of pattern (2) for a given pair. This effect is observable

in structures of the obtained solutions. Figs. 16 and 17 present a juxtaposition of solutions found for
CE

Nugent 16 by means of both a classic crisp algorithm and the simulated annealing based on linguistic

patterns with a five step granularity. By comparing these layouts, it can be effortlessly observed that the
AC

LP algorithm produced a solution where the full truth was present for all pairs with Truth_of(l) = 1. In

the layout from Fig. 16, which is optimal in crisp terms, four out of 11 pairs have locations worse than

optimal, when the Truth_of(pattern (2)) is concerned. However, on the other hand, the layout

demonstrated in Fig. 17 has a 6% worse function (1) value than the solution from Fig. 16.

23
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M

Fig. 16. An optimal solution for Nugent 16 on a 4×4 grid. Only maximal links between objects are displayed.
Objects number 2, 3, and 4 were slightly moved to show the appropriate relationships.
ED
PT
CE
AC

24
 ACCEPTED MANUSCRIPT

T
IP
CR
US
AN
M

Fig. 17. The best solution for Nugent 16 obtained by the proposed method for experts‘ opinions expressed on the
five steps scale. All links with the maximal Truth_of(l) = 1.
ED

4.2.2. Objects with variable sizes

The second experiment was conducted to check the usability of the proposed approach for problems
PT

involving items of variable sizes. Each object was modeled by an appropriate number of unitary

building blocks in such a way that their aggregate area was equal or slightly bigger than the object area
CE

provided in the problem description. For each object, one of its components was denoted as an input
AC

and output place. This item was linked with all other inner object building blocks with the relationship

value being twice as big as the largest link in the original relationship matrix. Connections between

input/output items of consecutive objects corresponded directly to the links between initial objects.

For comparisons, we took two data sets from literature, namely: a simple example examined by Jajodia

et al. (1992) and more sophisticated based on a real life data example analyzed by Tompkins (1978).

Both analyzed examples are similar in terms of the number of objects involved (7, 8) but they differ

25
 ACCEPTED MANUSCRIPT

significantly in the number of building blocks used for approximating areas covered by objects (from

one to three in the first, and from three to eight in the second example).

The experimental procedure started with arranging objects according to the best solutions reported in

referenced works and presented in Figs. A.1., and A.3 in Appendix. Tables A.1. and A.2. from

Appendix include link values and the number of components in every object. Since original

optimizations were carried out for distances between objects‘ centroids, therefore, we first optimized

T
locations of input/output building blocks within objects without changing their shapes. This was done

IP
by applying classic simulated annealing ten times. Results with the smallest transportation costs of this

CR
local optimization are demonstrated in the first column of Table 3. Next, the same crisp version of the

algorithm was used ten times for all building blocks in a given example. Objects‘ integrity was

US
controlled during this global optimization process. The best solutions are put into the second column

of Table 3.
AN

Table 3. Results of the second set of experiments for objects with variable sizes.
M

Characteristics of LP
Crisp data Comparisons
solutions for best layouts
Example Original layout Layout after Mean truth Mean Mean Crisp value
ED

after local opti- global optimi- upper truth truth for LP solu-
mization (Q1) zation (Q2) bound value corrected tion (Q3) [%] [%]
Jajodia et al. (1992)
154.5 153.2 .894 .888 .992 154.2 -.2 .6
Euclidean
PT

Jajodia et al. (1992)

165 157 .925 .917 .986 163 -1.2 3.8
Manhattan
Tompkins (1978)
423 356.5 .905 .891 .985 365.4 -15 1.9
CE

Euclidean
Tompkins (1978)
496 410 .933 .911 .976 423 -14.7 3.1
Manhattan
AC

Likewise in the previous set of experiments, it was assumed that the expert is rational and capable of

assessing relationship intensities on a five-step scale from Very Small to Very Big. The truth values of

increasing levels of links increase linearly. For increasing distances, the truth values decrease linearly in

such a way that truth value of the SMALL expression equals one for directly adjacent cells and zero for

items located at the ends of the layout area diagonals. Links between components of the same object

26
 ACCEPTED MANUSCRIPT

were set at the Very Big level. Under these assumptions, we ran the linguistic expression based simulated

annealing algorithm ten times.

From among the solutions where none of the objects was disintegrated, we selected layouts with the

highest mean truth. Parameters of the best results obtained according to the described procedure are

put together in Table 3. As in previous experiments, the quality of results was assessed by comparing

mean truths with maximal mean truths (upper bounds). These values are put together in columns three-

T
five of Table 3. They include upper bounds for the mean truth of pattern (2), mean truths of the best

IP
solution, and corrected mean truth for the best layout, respectively. The corrected mean truth was

CR
computed as a percentage difference between the upper bound and the highest mean truth. For the

best layouts in terms of the mean truth, we calculated the classic goal function (1) for crisp data

US
(column six of Table 3) which enabled comparisons between linguistic-based and crisp solutions -

columns seven and eight of Table 3.

AN
The comparison of the results presented in columns one and two indicates the superiority of the global

optimization over the local one, especially in the Tompkins‘ (1978) example. Since the mean truth value
M

is very close to one in both cases, the algorithm finds almost ideal solutions, that is, such layouts that
ED

meet fully pattern (2).

Columns seven and eight of Table 3 contain comparisons between results from columns one and two
PT

and the LP approach. The most reliable indicators are included in column eight of Table 3. They
CE

confront global optimization data (column two of Table 3) with the sixth column containing linguistic

pattern algorithm outcomes. The appropriateness of this comparison results from the fact that
AC

solutions presented in original works were obtained under different constraints and assumption from

ours. For instance, Tompkins (1978) provided results for regular rectangular objects while in the

present study, objects could take any shapes consisting of adjacent building blocks. In this example, the

LP methodology provided solutions differing from the traditional one by less than 2% (Euclidean) and

3.1% (Manhattan). For the Jajodia et al. (1992) problem the discrepancy amounted to barely .6%

(Euclidean) and 3.8% (Manhattan). What is interesting, in all trials, none of the objects was

27
 ACCEPTED MANUSCRIPT

disintegrated except for a single solution of the Tompkins‘ (1978) example, where one object fell apart.

However, this problem can be easily solved by manually rearranging two cells in a way demonstrated in

Fig. 18.

T
IP
CR
US
AN
M
ED

Fig. 18. One of the solutions for the Tompkins‘ (1978) example with one disintegrated object. Red arrows
indicate how to fix the problem.
Such a change produces a solution by 4% worse than the best outcome for crisp data. This situation
PT

suggests that decision makers may also consider solutions with some degree of objects‘ disintegration.

Thus, the schematic diagram of using the LP approach may look as in Fig. 19.
CE

Expert(s)’ knowledge
AC

Yes
Objects: Patterns: Algorithm: Execute Analyze Accept? Acceptable
layout solution
- Size - Expressions - Epoch
- Area type - Variables length No
- Distance - Links - Cooling
measure scheme Repair?
- Probability
No Yes

28
 ACCEPTED MANUSCRIPT

Fig. 19. A schematic diagram of using the LP approach for solving facility layout problems.

5. Conclusions
The proposed implementation of the simulated annealing algorithm with linguistic patterns is

promising. First of all, the procedure allows finding solutions with high mean truth for complex facility

layout problems. The presented simulation experiments show that optimizations based on linguistic

data, retrieved from a rational expert, provide solutions consistent to a high extent with the best crisp

T
layouts. The examination of the influence of algorithm parameters on the solution quality, yielded

IP
rather unexpected outcomes. Surprisingly, different values of the probability of accepting worse

CR
solutions (P) did not influence the mean truth of pattern (2). This preliminary result, leaning to

hypothesize that linguistic pattern based results are independent of the probability P requires, naturally,

US
further research and a confirmation for facility layout problems having other characteristics than these

examined in the current study. This direction of future investigations seems to be important as
AN
decreasing the P distinctly shortens calculation times. The effect of obtaining better mean results for a

slower cooling scheme (r = .99) also needs to be further analyzed. Applying this value, unfortunately,
M

markedly increase computation times. Though the time needed to complete the procedure was not
ED

subject to this study examination, a middle-class computer needed about 15 minutes to complete the

algorithm with r = .99 for the biggest 8×8 problem. Therefore, the proposed here method with the
PT

optimal high value of the cooling scheme parameter can be in practice applied, at most, to medium-

sized facility layout problems. On the other hand, such problem sizes are prevailing in the design
CE

practice.
AC

The outcomes involving objects with unequal dimensions indicate an interesting area of further studies.

The presented concept of modeling transportation systems within complex objects provided solutions

in which the building blocks fall apart rarely and generate concentrated structures for the open field

model. Thus, it may be interesting to examine other concepts of intracellular transportation models

(e.g., loop) and their influence on objects‘ shapes while applying the proposed approach.

29
 ACCEPTED MANUSCRIPT

The presented simulation results involve one of the easiest possible applications of the LP concept with

only a single pattern used as a criterion and with the assumption of linear relationships between truth

assessments of consecutive levels of linguistic expressions. The software developed for conducting

current experiments could be redesigned to take into account multiple criteria and various ways of

defining linguistic expressions and patterns. However, in the case of multiple criteria, it would be very

difficult to use the same as in the present study idea involving comparisons of the LP approach with

a hypothetical objective system. From one hand side, the mean truth upper bound could not be directly

T
IP
applied and, from the other, it might be problematic to find clear crisp counterparts of subjectively

defined variables and patterns. Undoubtedly, developing an appropriate research methodology in this

CR
regard will constitute a challenging direction of further investigations.

Acknowledgements
US
The work was partially financially supported by the Polish National Science Center grant no.
2011/03/B/HS4/04532.
AN
References
1. Aiello, G., & Enea, M. (2001). Fuzzy approach to the robust facility layout in uncertain
M

production environments. International Journal of Production Research, 39(18), 4089–4101.

doi: 10.1080/00207540110061643
2. Aiello, G., La Scalia, G., & Enea, M. (2012). A multi objective genetic algorithm for the facility
ED

layout problem based upon slicing structure encoding. Expert Systems with Applications,
39(12), 10352–10358. doi: 10.1016/j.eswa.2012.01.125
3. Aiello, G., La Scalia, G., & Enea, M. (2013). A non dominated ranking Multi Objective Genetic
PT

Algorithm and ELECTRE method for unequal area facility layout problems. Expert Systems
with Applications, 40(12), 4812–4819. doi: 10.1016/j.eswa.2013.02.026
4. Altuntas, S., Selim, H., & Dereli, T. (2014). A fuzzy DEMATEL-based solution approach for
CE

facility layout problem: a case study. The International Journal of Advanced Manufacturing
Technology, 73(5-8), 749–771. doi: 10.1007/s00170-014-5826-3
5. Azadeh, A., Moghaddam, M., Nazari, T., & Sheikhalishahi, M. (2016). Optimization of facility
AC

layout design with ambiguity by an efficient fuzzy multivariate approach. The International
Journal of Advanced Manufacturing Technology, 84(1–4), 565–579. doi:10.1007/s00170-015-
7714-x
6. Babbar-Sebens, M., & Altuntas, B. S. (2012). Interactive Genetic Algorithm with Mixed
Initiative Interaction for multi-criteria ground water monitoring design. Applied Soft
Computing, 12(1), 182–195. doi: 10.1016/j.asoc.2011.08.054
7. Bashiri, M., & Hosseininezhad, S. J. (2008). A fuzzy group decision support system for
multifacility location problems. The International Journal of Advanced Manufacturing
Technology, 42(5–6), 533–543. doi: 10.1007/s00170-008-1621-3

30
 ACCEPTED MANUSCRIPT

8. Cheng, R., Gent, M., & Tozawa, T. (1995). Genetic search for facility layout design under
interflows uncertainty. In, IEEE International Conference on Evolutionary Computation, 1995
(V. 1, s. 400-). doi: 10.1109/ICEC.1995.489181
9. Deb, S. K., & Bhattacharyya, B. (2005). Fuzzy decision support system for manufacturing
facilities layout planning. Decision Support Systems, 40(2), 305–314. doi:
10.1016/j.dss.2003.12.007
10. Drira, A., Pierreval, H., & Hajri-Gabouj, S. (2007). Facility layout problems: A survey. Annual
Reviews in Control, 31(2), 255–267. doi: 10.1016/j.arcontrol.2007.04.001
11. Dweiri, F., & Meier, F. A. (1996). Application of fuzzy decision-making in facilities layout
planning. International Journal of Production Research, 34(11), 3207–3225. doi:
10.1080/00207549608905085

T
12. Evans, G. W., Wilhelm, M. R., & Karwowski, W. (1987). A layout design heuristic employing
the theory of fuzzy sets. International Journal of Production Research, 25(10), 1431–1450. doi:

IP
10.1080/00207548708919924
13. García-Hernández, L., Arauzo-Azofra, A., Salas-Morera, L., Pierreval, H., & Corchado, E.

CR
(2015a). Facility layout design using a multi-objective interactive genetic algorithm to support
the DM. Expert Systems, 32(1), 94–107. doi: 10.1111/exsy.12064
14. García-Hernández, L., Palomo-Romero, J. M., Salas-Morera, L., Arauzo-Azofra, A., &

US
Pierreval, H. (2015b). A novel hybrid evolutionary approach for capturing decision maker
knowledge into the unequal area facility layout problem. Expert Systems with Applications,
42(10), 4697–4708. doi: 10.1016/j.eswa.2015.01.037
AN
15. García-Hernández, L., Pierreval, H., Salas-Morera, L., & Arauzo-Azofra, A. (2013). Handling
qualitative aspects in Unequal Area Facility Layout Problem: An Interactive Genetic Algorithm.
Applied Soft Computing, 13(4), 1718–1727. doi: 10.1016/j.asoc.2013.01.003
M

16. Gonçalves, J. F., & Resende, M. G. C. (2015). A biased random-key genetic algorithm for the
unequal area facility layout problem. European Journal of Operational Research, 246(1), 86–
107. doi: 10.1016/j.ejor.2015.04.029
ED

17. Grobelny, J. (1987a). The fuzzy approach to facilities layout problems. Fuzzy Sets and Systems,
23(2), 175–190. doi: 10.1016/0165-0114(87)90057-1
18. Grobelny, J. (1987b). On one possible „fuzzy‖ approach to facilities layout problems.
PT

International Journal of Production Research, 25(8), 1123.

19. Grobelny, J. (1988). The ‗linguistic pattern‘ method for a workstation layout analysis.
International Journal of Production Research, 26(11), 1779–1798. doi:
CE

10.1080/00207548808947991
20. Grobelny, J. (2016). Fuzzy-based linguistic patterns as a tool for the flexible assessment of
a priority vector obtained by pairwise comparisons. Fuzzy Sets and Systems. doi:
AC

10.1016/j.fss.2015.05.012
21. Guan, J., & Lin, G. (2016). Hybridizing variable neighborhood search with ant colony
optimization for solving the single row facility layout problem. European Journal of
Operational Research, 248(3), 899–909. doi: 10.1016/j.ejor.2015.08.014
22. Haktanirlar Ulutas, B., & Kulturel-Konak, S. (2012). An artificial immune system based
algorithm to solve unequal area facility layout problem. Expert Systems with Applications,
39(5), 5384–5395. doi: 10.1016/j.eswa.2011.11.046
23. Heragu, S. S., & Alfa, A. S. (1992). Experimental analysis of simulated annealing based
algorithms for the layout problem. European Journal of Operational Research, 57(2), 190–202.
doi: 10.1016/0377-2217(92)90042-8

31
 ACCEPTED MANUSCRIPT

24. Jajodia, S., Minis, I., Harhalakis, G., & Proth, J.-M. (1992). CLASS: Computerized LAyout
Solutions using Simulated annealing. International Journal of Production Research, 30(1), 95–
108. doi:10.1080/00207549208942880
25. Karray, F., Zaneldin, E., Hegazy, T., Shabeeb, A. H. M., & Elbeltagi, E. (2000). Tools of soft
computing as applied to the problem of facilities layout planning. IEEE Transactions on Fuzzy
Systems, 8(4), 367–379. doi: 10.1109/91.868944
26. Kirkpatrick, S., Gelatt, C. D., & Vecchi, M. P. (1983). Optimization by Simulated Annealing.
Science, 220(4598), 671–680. doi:10.1126/science.220.4598.671
27. Komarudin, & Wong, K. Y. (2010). Applying Ant System for solving Unequal Area Facility
Layout Problems. European Journal of Operational Research, 202(3), 730–746. doi:
10.1016/j.ejor.2009.06.016
28. Kumar, R. M. S., Asokan, P., & Kumanan, S. (2008). Artificial immune system-based algorithm

T
for the unidirectional loop layout problem in a flexible manufacturing system. The International

IP
Journal of Advanced Manufacturing Technology, 40(5-6), 553–565. doi: 10.1007/s00170-008-
1375-y

CR
29. Kundu, A., & Dan, P. K. (2012). Metaheuristic in facility layout problems: current trend and
future direction. International Journal of Industrial and Systems Engineering, 10(2), 238–253.
doi: 10.1504/IJISE.2012.045182

US
30. Kusiak, A., & Heragu, S. S. (1987). The facility layout problem. European Journal of
Operational Research, 29(3), 229–251. doi: 10.1016/0377-2217(87)90238-4
31. Liggett, R. S. (1981). The Quadratic Assignment Problem: An Experimental Evaluation of
AN
Solution Strategies. Management Science, 27(4), 442–458. doi: 10.1287/mnsc.27.4.442
32. Liggett, R. S. (2000). Automated facilities layout: past, present and future. Automation in
Construction, 9(2), 197–215. doi: 10.1016/S0926-5805(99)00005-9
M

33. Lim, Z. Y., Ponnambalam, S. G., & Izui, K. (2017). Multi-objective hybrid algorithms for layout
optimization in multi-robot cellular manufacturing systems. Knowledge-Based Systems. doi:
10.1016/j.knosys.2016.12.026
ED

34. Loiola, E. M., de Abreu, N. M. M., Boaventura-Netto, P. O., Hahn, P., & Querido, T. (2007). A
survey for the quadratic assignment problem. European Journal of Operational Research,
176(2), 657–690. doi: 10.1016/j.ejor.2005.09.032
PT

35. Matai, R. (2015). Solving multi objective facility layout problem by modified simulated
annealing. Applied Mathematics and Computation, 261, 302–311. doi:
10.1016/j.amc.2015.03.107
CE

36. Mohamadghasemi, A., & Hadi-Vencheh, A. (2012). An Integrated Synthetic Value of Fuzzy
Judgments and Nonlinear Programming Methodology for Ranking the Facility Layout Patterns.
Comput. Ind. Eng., 62(1), 342–348. doi: 10.1016/j.cie.2011.10.004
AC

37. Moslemipour, G., Lee, T. S., & Rilling, D. (2011). A review of intelligent approaches for
designing dynamic and robust layouts in flexible manufacturing systems. The International
Journal of Advanced Manufacturing Technology, 60(1-4), 11–27. doi: 10.1007/s00170-011-
3614-x
38. Muther, R., (1961), Systematical Layout Planning (Boston, MA: Industrial Education Institute).
39. Nugent, C. E., Vollmann, T. E., & Ruml, J. (1968). An Experimental Comparison of
Techniques for the Assignment of Facilities to Locations. Operations Research, 16(1), 150–173.
doi:10.1287/opre.16.1.150
40. Ohmori, S., Yoshimoto, K., & Ogawa, K. (2010). Solving Facility Layout Problem via Particle
Swarm Optimization. In: 2010 Third International Joint Conference on Computational Science
and Optimization (CSO) (V. 1, pp. 409–413). doi: 10.1109/CSO.2010.242

32
 ACCEPTED MANUSCRIPT

41. Önüt, S., Tuzkaya, U. R., & Doğaç, B. (2008). A particle swarm optimization algorithm for the
multiple-level warehouse layout design problem. Computers & Industrial Engineering, 54(4),
783–799. doi: 10.1016/j.cie.2007.10.012
42. Paes, F. G., Pessoa, A. A., & Vidal, T. (2017). A hybrid genetic algorithm with decomposition
phases for the Unequal Area Facility Layout Problem. European Journal of Operational
Research, 256(3), 742–756. doi: 10.1016/j.ejor.2016.07.022
43. Palomo-Romero, J. M., Salas-Morera, L., & García-Hernández, L. (2017). An island model
genetic algorithm for unequal area facility layout problems. Expert Systems with Applications,
68, 151–162. doi: 10.1016/j.eswa.2016.10.004
44. Palubeckis, G. (2015). Fast simulated annealing for single-row equidistant facility layout.
Applied Mathematics and Computation, 263, 287–301. doi: 10.1016/j.amc.2015.04.073

T
45. Parsaei, H. R., & Galbiati III, L. J. (1987). Facilities planning and design with microcomputers.
Computers & Industrial Engineering, 13(1–4), 332–335. doi:10.1016/0360-8352(87)90109-4

IP
46. Raoot, A. D., & Rakshit, A. (1991). A ‗fuzzy‘ approach to facilities lay-out planning.
International Journal of Production Research, 29(4), 835–857. doi:

CR
10.1080/00207549108930105
47. Raoot, A. D., & Rakshit, A. (1993). A ‗linguistic pattern‘ approach for multiple criteria facility
layout problems. International Journal of Production Research, 31(1), 203–222. doi:
10.1080/00207549308956721

Computers &
US
48. Sadrzadeh, A. (2012). A genetic algorithm with the heuristic procedure to solve the multi-line
layout problem. Industrial Engineering, 62(4), 1055–1064.
AN
doi:10.1016/j.cie.2011.12.033
49. Samarghandi, H., Taabayan, P., & Jahantigh, F. F. (2010). A particle swarm optimization for the
single row facility layout problem. Computers & Industrial Engineering, 58(4), 529–534. doi:
10.1016/j.cie.2009.11.015
M

50. Sahni, S., & Gonzalez, T. (1976). P-Complete Approximation Problems. Journal of the ACM,
23(3), 555–565. doi: 10.1145/321958.321975
ED

51. Sangwan, K. S. (2010). Fuzziness in Materials Flow and Plant Layout. In: C. Kahraman & M.
Yavuz (Ed.), Production Engineering and Management under Fuzziness (pp. 359–380).
Springer Berlin Heidelberg. doi: 10.1007/978-3-642-12052-7_15
PT

52. Sargent, T. A., Kay, M. G., & Sargent, R. G. (1997). A Methodology for Optimally Designing
Console Panels for Use by a Single Operator. Human Factors: The Journal of the Human
Factors and Ergonomics Society, 39(3), 389–409. doi: 10.1518/001872097778827052
CE

53. Seehof, J. M., Evans, W. O., Friederichs, J. W., & Quigley, J. J. (1966). Automated Facilities
Layout Programs. In: Proceedings of the 1966 21st National Conference (pp. 191–199). New
York, NY, USA: ACM. doi: 10.1145/800256.810696
AC

54. Skorin-Kapov, J. (1994). Extensions of a tabu search adaptation to the quadratic assignment
problem. Computers & Operations Research, 21(8), 855–865. doi: 10.1016/0305-
0548(94)90015-9
55. Taillard, E. (1991). Robust taboo search for the quadratic assignment problem. Parallel
Computing, 17(4–5), 443–455. doi: 10.1016/S0167-8191(05)80147-4
56. Tompkins, J.A. (1978). How to gather the data you need. Modern Materials Handling, June, 50-
56.
57. Tompkins, J.A., White, J.A.A. (1984). Facilities Planning, New York: John Wiley & Sons.
58. Wilhelm, M. R., Ward, T. L. (1987). Solving quadratic assignment problems by ‗Simulated
Annealing‘. IIE Transactions, 19(1), 107–119. doi:10.1080/07408178708975376

33
 ACCEPTED MANUSCRIPT

59. Zadeh, L.A. (1978a). Fuzzy sets as a basis for a theory of possibility, Fuzzy sets and Systems,
1(1), 3-28. doi: 10.1016/0165-0114(78)90029-5
60. Zadeh, L.A. (1978b). PRUF—a meaning representation language for natural languages.
International Journal of Man-Machine Studies, 10(4), 395–460. doi: 10.1016/S0020-
7373(78)80003-0
61. Zhao, F., Li, G., Yang, C., Abraham, A., & Liu, H. (2014). A human computer cooperative
particle swarm optimization based immune algorithm for layout design. Neurocomputing,
132(0), 68–78. doi: 10.1016/j.neucom.2013.03.062

T
IP
CR
US
AN
M
ED
PT
CE
AC

34
 ACCEPTED MANUSCRIPT

Appendix
Data for Jajoda et al. (1992) and Tompkins (1978) examples with variable objects‘ sizes.

Example from Jajodia et al. (1992)

The links between building blocks of the same object were specified as twice as big as the maximum
value from the objects‘ relationships matrix, that is 18.
Table A.1. Links between objects and the number of building blocks for every object in the Jajodia et al. (1992)
example.
Object Number
1 2 3 4 5 6 7
number of blocks
1 3 0

T
2 2 9 0

IP
3 2 6 4 0
4 2 0 4 0 0

CR
5 1 0 4 0 4 0
6 1 0 4 4 0 4 0
7 1 3 0 0 4 0 0 0

US
AN
M
ED
PT

Fig. A.1. The original best solution reported in Jajodia et al. (1992).
CE
AC

Fig. A.2. The best solution obtained by means of our LP approach for Jajodia et al. (1992) example.

35
 ACCEPTED MANUSCRIPT

Example from Tompkins (1978)

In this example, the area of a building block was set at 1000 units and RelCharts weight values were set
at A=4, E=3, I=2, O=1. The links between building blocks of the same object were specified as twice
as big as the maximum value from the objects‘ relationships matrix, that is eight.

Table A.2. Links between objects and the number of building blocks for every object in the Tompkins (1978)
example.
Object Area
1 2 3 4 5 6 7 8
number (blocks)
1 7500 (8) ×
2 7000 (7) 2 ×

T
3 3000 (3) 0 2 ×

IP
4 5000 (5) 1 1 0 ×
5 4500 (5) 0 0 0 2 ×

CR
6 3000 (3) 1 2 1 2 3 ×
7 7000 (7) 0 0 0 0 0 3 ×
8 3000 (3) 2 0 0 0 0 0 4 ×

US
AN
M
ED
PT

Fig. A.3. The original best solution for centroids reported in Tompkins (1978). 00 denotes an empty cell.
CE
AC

Fig. A.4. The best solution obtained by means of our LP approach for Tompkins (1978) example. 00 denotes an
empty cell.

36